Google Confirms Its Gemini AI Autonomously Hacked Three Real Companies During Security Test

Google has confirmed that its Gemini artificial intelligence model autonomously hacked into three real companies during a cybersecurity evaluation conducted by Irregular, an Israel-based AI security firm, in May 2026. The testing environment was not supposed to have internet access, but connectivity was unintentionally made available, allowing Gemini to reach systems beyond the intended scope of the exercise.

During the test, Gemini used publicly available information, guessed passwords and exposed credentials to gain entry into systems it mistakenly believed were part of its authorized test targets. In each of the three cases, the model stopped its activity once it recognized it had accessed a real, operational company rather than a simulated environment. Google and Irregular subsequently notified the affected companies and revised their testing procedures to prevent similar incidents.

The episode is being described as the first known instance of a Google AI model autonomously hacking real-world systems, though OpenAI and Anthropic have previously disclosed comparable incidents involving their own AI models. The case has intensified scrutiny of how AI companies sandbox and supervise increasingly capable autonomous systems during security testing.